A heat flow for special metrics Hartmut Weifi and Frederik Witt September 6, 2011 Abstract On the space of positive 3-forms on a seven-manifold, we study a natural functional whose critical points induce metrics with holonomy contained in G2- We prove short-time existence and uniqueness for its negative gradient flow. Furthermore, we show that the flow exists for all times and converges modulo diffeomorphisms to some critical point for any initial condition sufficiently C°°-close to a critical point. MSC 2000: 53C44 (35G25, 35K55, 53C25) KEYWORDS: G2-manifolds, geometric evolution equations, quasilinear parabolic equations 1 Introduction A central problem in Riemannian geometry is the construction of metrics with prescribed properties of the Ricci tensor. In this context the metrics we refer to as special are partic- ularly interesting. These metrics are induced by a differential form Q of special algebraic type subject to the non-linear harmonic equation dn = o, d* n n = 0, (1) where *q is the Hodge operator associated with the induced metric g^. A prototypical example are so-called G2-metrics. A theorem of Fernandez and Gray [TT] asserts <Q to hold if and only if the holonomy of g^ is contained in G2- This, in turn, implies gci to be Ricci flat [3]- Other metrics of this type are Spin(7)-metrics (which are also Ricci flat) or quaternionic Kahler metrics (which are Einstein), but more exotic examples such as PSU(3)-metrics |14] (satisfying a less standard condition on the Ricci tensor) also fit into this setting. In the G2-case we consider positive 3-forms O over an oriented 7-manifold which are special in as far they give rise to a complementary 4-form 8(0) so that the volume form vol^i : = O A 0(0)/7 induces the chosen orientation. In fact, once the metric has been constructed from O one has 0(0) = *qO. If M is compact, ([I]) is equivalent to the non-linear Laplace equation A^O = 0, where A^ is the Hodge Laplacian associated with g^. In this sense, O is "self-harmonic", but we shall stick to the usual G2~j argon and refer to positive forms satisfying ([I]) as torsion-free. For closed O, Hitchin |14| interpreted the second condition as the Euler-Lagrange equation for the functional on positive 3-forms O i— > j M vola restricted to the cohomology class [O]. Existence of critical points, however, is a delicate issue. Since Joyce's seminal work |16] we know non-trivial compact holonomy G2-manifolds to exist, but a Yau-Aubin type theorem which guarantees a priori existence is yet missing. 1 A natural idea for proving existence of G2-metrics is to look for a geometric evolution equation on the space of positive 3-forms, which could evolve forms towards a torsion-free f2, cf. [T7] for a rather general perspective. A first candidate for a flow equation has been proposed by Bryant in [5], namely the "Laplacian flow" Restricted to closed positive 3-forms, we can think of this flow as the (L 2 -)gradient flow of Hitchin's functional. However, as we are going to show, the resulting flow equation is not even weakly parabolic so that standard parabolic theory does not apply directly (though Bryant and Xu, resp. Xu and Ye have managed to establish short-time existence for the Laplacian flow for closed initial conditions, cf. [SJ|25]). This is reminiscent of the Einstein- Hilbert functional whose negative gradient is difficult to deal with on the same grounds, a fact which subsequently led to the definition of Ricci flow. We therefore consider the negative gradient flow of the "Dirichlet energy" functional whose critical points, as we will show, are precisely given by torsion-free forms. In principle, the definition of T> makes sense for any special metric. The reason to focus on G2 is twofold. The set of positive 3-forms Q is an open subset of il 3 (M), and G2 acts transitively on the sphere (as do all reduced holonomy groups of manifolds which are not locally symmetric). Both these features greatly simplify technicalities. Our first result is short-time existence and uniqueness. Theorem. Let M be a closed, oriented 7-manifold. Given a positive 3-form f^o on M, there exists e > and a smooth family of positive 3-forms Q(t) for t € [0, e] such that Furthermore, for any two solutions Q(t) and Q'(t) we have fi(t) = Q'(t) whenever defined. Hence we can speak of the Dirichet energy flow for some initial condition S}q defined on a maximal time— interval [0, T max ), < T ma x 00. We emphasise that for the Dirichlet energy flow the initial condition does not have to be assumed to be closed. The proof is based on DeTurck's trick as introduced in [8], namely to consider a geometric perturbation Q of the negative gradient of T>. Then Q is strongly elliptic and the standard theory of quasilinear parabolic equations applies. Of course, one cannot expect longtime existence and convergence in general as the existence of G2-metrics is topologically obstructed. For instance, if M is compact then O cannot be exact, hence the third Betti number 63 must be greater or equal to 1. However, a meaningful negative gradient flow is certainly expected to exist for all times and to converge near a critical point. Indeed we will prove the subsequent stability result (cf. Theorem 18.11 and Corollary 18.21 for the precise statement). Theorem. Let M be a closed, oriented 7-manifold and let Cl be a torsion-free G2~form on M. For initial conditions sufficiently C°°-close to f2 the Dirichlet energy flow exists for all times and converges modulo diffeomorphisms to a torsion-free G2-form. — n = A n n. — = -gradP(fi), n(0) = n . 2 The theorem resembles Sesum's corresponding stability result for Ricci flat metrics under Ricci flow [21]. Nevertheless, two major differences occur: For stability of the Ricci flow one needs to assume that (a) the Ricci flat metric g (which corresponds to the torsion- free O in our setting) is a smooth point of the moduli space of Ricci flat metrics (b) the linearisation at g of the quasilinear elliptic operator involved (corresponding to our Q) is non-positive. Both conditions are difficult to check in practice unless one makes further assumptions such as special holonomy. In our situation, however, these assumptions are automatically satisfied. In particular, we can show that near J7, the set Q~ 1 (0) provides a slice for the action of diffeomorphisms isotopic to the identity on the space of torsion-free G2-metrics. Consequently, the moduli space of torsion-free G2~forms is smooth. This has been proven previously by Joyce (see [IB]), but our approach is rather different and based on the specific geometry of the zero sets of grad£> and Q. Secondly, we note that on a technical level, the linearisation Lq of at f2 does not admit a Weitzenbock formula as does the Lichnerowicz Laplacian which appears in the Ricci flow case. However, we can prove the Garding type inequality where Cl G fl 3 (M) and C is some positive constant. With this coercivity condition we can invoke a result of Lax-Milgram type in order to establish longtime existence following ideas developed in |15| . Finally, convergence follows from a careful analysis of the remainder term Rq = Q — Lq. The flow is thus, in principle, capable of finding torsion-free G2-forms on a 7-manifold if they exist. 2 G2— structures Here and in the sequel, we let M be a connected, closed, oriented manifold of dimension 7. We first recall some basic features of G2-geometry to fix notations. Good references are [3] and Chapter 10 in [16] , The group GL(7) acts on A 3 R 7 * and has an open orbit O diffeomorphic to GL(7)/G2- In fact, O is an open cone, for c • Id 6 GL(7), c 7^ acts on any 3-form by multiplication with c~ 3 . Since G2 is a subgroup of SO(7), any Q G O induces an orientation and a Euclidean metric g^ on M 7 . We denote by A+ those which induce the standard orientation on ]R 7 and refer to its elements as positive forms. They are acted on transitively by GL(7)+, the orientation-preserving linear isomorphisms of R 7 . Let tt%(M) denote the open set of sections of A+M, the fibre bundle associated with Aj_. Then a section G Q^_(M) (which exists if and only if the second Stiefel- Whitney class of M vanishes) induces a reduction of the frame bundle to a principal G2-bundle. We also refer to the pair (M, f2) as a Q2-structure. Such a structure singles out a principal SO (7)- bundle whose associated metric, Hodge star operator and Levi-Civita connection we denote by gQ, *q and respectively. Locally, there exist so-called G2-fram.es, i.e. local frames (ei, . . . , ej) of TM for which Q has "normal form". In our convention, we can write O with respect to a G2-frame as n = e 127 + e 347 + e 567 + e 135 _ e 14 6 _ e 236 _ e 245 with e lJ ' fc shorthand for e* A e J A e k . Note that a G2-frame is orthonormal for the induced G 2 -metric g n . 3 The holonomy of gn is contained in G2 (implying that gn is Ricci-flat [3]) if and only if the underlying G2-form O is parallel, i.e. V sn J7 = 0. In this case we shall say that the G2-structure is torsion-free while we call (M, f2) a holonomy G2 -manifold if the holonomy of g$i is actually equal to G2- (This convention is by no means universal in the literature.) A torsion-free G2~structure has holonomy G2 if the fundamental group 7Ti(M) is finite (M being compact). By a theorem of Fernandez and Gray torsion-freeness is equivalent to dVl = and 5$fl = 0, where 5q = ( — l) p *nd*n is the induced codifferential on p-forms. We shall therefore refer to any such as a torsion-free G2-form. The latter equation can be viewed as the Euler-Lagrange equation of a non-linear variational problem set up by Hitchin |14] , Consider the smooth GL(7)+-equivariant map cj) : A 3 — > A 7 , Q 1 — y voln := *nl = lQ A *nf2, whose first derivative at Q evaluated on Cl G A 3 is £>n>(f2) = J*nfiAf2. (2) Integrating (p gives the functional n ■. n%{M) —?■ e, n ^ I <p{n). (3) JM In analogy with Hodge theory we can restrict % to a fixed cohomology class and ask for critical points. From (|2|) it follows that a closed £1 is a critical point in its cohomology class if and only if 5ffl = [J3]. In particular, is torsion-free and thus harmonic with respect to its induced Laplacian = d5n + fed. In passing we note that modulo a constant, %{Q) can be viewed as the norm square of the Euler vector field $7 i— > on the "prehilbert" manifold Q^_(M) with induced L 2 -metric (Oi,r2 2 ) L 2 := / gn{&i,&2)vol n = / OiA*n^2, JM JM for elements Oi, in the tangent space TqQ^M) = n 3 (M). We will drop the reference to £1 whenever this can be safely done and simply write (• , -) L 2 and g. The associated norms are then denoted by || • || and | • | respectively. 3 Representation theory Next we recall some elements of G2~representation theory. Most of the material is standard (mainly taken from [5], [7] and |18j ) or follows from straightforward computations. The group G2 acts irreducibly in its vector representation A 1 = M. 7 (in presence of a metric, we tacitly identify vectors with their duals). This action extends to the exterior algebra in the standard fashion, though A p , the G2~representation over p-forms, is no longer irreducible for 2 < p < 5. More precisely, we have orthogonal decompositions A 2 = A? © A 2 4 , A 3 = A? © Af © A^ 7 , where the subscript indicates the dimension of the module. We denote the corresponding components by [a p ] q . For p = 2,3 they satisfy A 2 = {a e A 2 I *n (a A H) = 2a}, Af 4 = {a G A 2 | *q (aAO) = -a}, (4) A3 = {* n (x A U) I X G A 1 }, Af 7 = {a G A 3 | (* n Sl) A a = 0, Q A a = 0}. 4 The Lie algebra of 02 sitting inside so (7) = A 2 corresponds to A 2 4 , while A 3 simply consists of multiples of Q. Note that by equivariance, *q induces isomorphisms Aq = A 7 q ~ p from which an analogous decomposition of A 4 and A 5 follows. This and the characterisations (|4|) are obtained from a routine application of Schur's lemma. For illustration, we derive for 7] G A 2 the identity (?7l0)l0 = 3[t?] 7 . (5) Here, i_ denotes the extension of the metric contraction to i_: A k V* (g> A l V* — > A l ~ k V*, e.g. e 12 Le 12345 = e 345 etc. Now r] i— > r]iSl is a G2-equivariant map taking values in the irreducible module A 1 = A 7 so that by Schur A 2 4 C keri_fi, whence = [rj\ 7 \St. Therefore, the identity ([5]) needs only to be checked for one nontrivial element in A 2 (again by Schur). Fixing a G2-frame as in the previous section, we find e\iSl = e27 + 635 — G A 2 , hence (eii_f2)i_Q = 3ei. In the same vein, we can prove (*n(aAO))AO = -4*oa, (★q (a A*nO)) A*qU = 3*na, (6) ( *q (a A *n^)) A = 2aA* n n. If the manifold M is endowed with a G2~structure f2, all these decompositions and iden- tities acquire global meaning. In particular we can speak of fig-forms, where fig(M) = C°°(AqT*M) are smooth sections of the bundles with fibre Aq. As in the case for Kahler manifolds, this decomposition gives rise to G2~analogues of the Cauchy-Riemann operator, provided the G2-structure is torsion-free. The subsequent formulas were derived by Bryant and Harvey and can be found in |5]. We briefly describe their results. First, we fix reference modules for the irreducible G2-representations occuring in the exterior algebra A*, namely fil = 0?(M), n 7 = n^(M), n u = ^14 (M) and n 2 7 = nij 7 (M). An y form « G n p (M) can be written in terms of G 2 -equivariant maps applied to elements of these reference modules. For instance Cl = [f2]i © [fi] 7 © [f2] 27 G fi 3 (M) can be written as Cl = ftt © -k n (a A fi) ffi 7 for / G Q\, a € ^7 and 7 G ^27- There exist first order differential operators d p : fi p — > £l q such that the identities of Table [1] hold. df = d\f d(fn) = 4/Afi d(f* n n) = d\f A*fi <ia = i *q A +d 7 u a disifaA*^) = — |dja • fi — ^ *n (4 a A f2) +dl 7 a d*i)(aA(l) = |d£a • +^4a A fi + *q d^a rf(aAO) = ^dja/\*nQ — -kn d\ 4 a d(a A *^fi) *n d 7 7 ot d(*na) = —d\a ■ voIq, d/3 = |* Q (4 4 /3Afi) +d 44 /3 d-y = \d 2 7 l A n +*n df 7 >y d(*nl) = -Idf'jA-knn -*nd?47 Table 1: Exterior derivative formulas 5 That such a table must exist follows from the torsion-freeness which is equivalent to finding coordinates xi,...,x 7 with d Xi £l(x) = around any x £ M. Put differently, the G2- structure (M, f2) locally osculates to first order to the flat structure on M. 7 . Therefore, the exterior derivative of expressions such as *o(aAfl) only depends on the 1— jet of the reference forms involved, e.g. a. The operators dq are obtained by compounding d with suitable G2- equivariant maps. As an example, da = *n(ct A © /3 for a € ^7 and /3 £ f^u, where a = *n(da A*n^)/3 and (3 = [da] 14. Working out the identities of Table [1] is then a matter of computation using the algebraic formulas (|6|) , (jl6[) and (|17p . Our precise definition of the operators dq can be found in Appendix lAl Remark: The operators dq and dp are formally adjoint to each other, i.e. for any a p £ fl p and a q £ il q . Example: With Table[T]at hand we compute the (co-)differential of Cl = /r2©*n(dAr2)©7 £ n 3 (M) and find dtt = fdjd *n n (4/ + \d 7 7 a + jdfj) Afie*o (d 7 27 a + 4^7) S n n = *n ((-4/ - |4" + ^4 7 7) A © ^4" + d?47- ( 7 ) Using Table [IJ <i 2 = implies the following second-order identities of Table [2] 44 = 444 = d 7 d 7 7 = 4 4 44 = 144 444 + 2 4I47 = 34X4 + 4 7 4 = d\df = 44 4 + 24 7 4f = 4 7 4 4 + 44^4 4 = 34 4 4i + 44 7 = 4 4 4 7 + 4di 7 df 7 = 24 7 4 7 - 44 7 = Table 2: Second order identities We will also need the Laplacians A^a p , a p £ O p . These are given in Table [3] An/ = 44/ A n a = [d 7 7 d 7 7 + d\d\)a An/3 = (|444 4 + 4I47)/3 An 7 = (^ 7 7 4 7 + 4 7 4J + (4 7 ) 2 )7 Table 3: Laplacians Example: If Q = fQ © *n(d A 0) © 7, then A n ^ = A n / • ^ © *n(And A 0) © A n7 . We finish this section with some material on SU(3)-representation theory. This will become relevant later, when we compute the symbols of various differential operators on a manifold 6 equipped with some fixed G2~structure. The finer SU(3)-representation theory will enable us to derive properties of these symbols such as invertibility and definiteness. We pick a unit vector £ 6 A 1 in the vector representation of G2- Since the unit sphere S 6 is diffeomorphic with G2/SU(3), £ gives rise to an SU(3)-representation over £ , namely the real representation underlying the complex vector representation C 3 . In particular, £^ carries a complex structure. In terms of forms, the group SU(3) can be regarded as the stabiliser of a non-degenerate 2-form u G A 2 £^ and a complex volume form ^ = tp + + i?p_ G A 3,0 ^. These forms relate to SI and *qS7 via !] = wA(|i/) + , (8) 1 9 * n n = V-A£ + -uA In terms of a G2-frame with £ = e 7 we find to = e 12 + e 34 + e 56 , ip + = e 135 — e 146 - e 236 - e 245 and tp- = e 136 + e 145 + e 235 — e 246 . They satisfy the algebraic relations uj A^ ± =0, ip+ A ip- = |w 3 . (9) The decomposition of the exterior algebra over into irreducibles is given by A 1 = A 2 = A 2 © A 2 , © A 2 ;, A 3 = A? + ©A?_©A^©A? 2 , (10) where A* := A 1 ^. As above the numerical subscript keeps track of the dimension. We also use these subscripts to denote the corresponding components of a form, e.g. 7 € A 3 can be decomposed into the direct sum 7 = 714. © 7i_ © 76 © 712- The two trivial representations A 3 _|_ are spanned by ip + and respectively, while A| corresponds to the Lie algebra of su(3) sitting inside so (6) = A 2 . More importantly for our purposes we can consider the decomposition of the exterior algebra over R into SU(3)-irreducibles. Here, we shall denote by (n)g the n-dimensional irreducible SU(3)-representation inside Aq. Then A 1 (I)* © (6) 7 , A 2 =S (l) 2 © (6) 2 7 © (6) 2 4 © {8)1 14) A 3 (l) 3 © (l) 3 © (6)? © (1) 3 7 © (6) 3 7 © (8) 3 7 © (12) 27- so that no confusion shall occur. The decomposition of A 3 is of particular importance for the sequel. The occuring modules can be characterised as follows: 1)? = {a{uo A£ + V+) 1 a G R}, 1)? = {bip- b G R}, l)Sr = {c(-4w A £ + 3^+) c G R}, 6) 3 = {(Xl<0_) A £ + (Xlw) A u 1 X G e L }, e)i 7 = {(FlV»_) a ^ - (Flw) a w 1 y e }• 8)27 = {/3 8 A £ | /3 8 G A 2 }. 1— > A(Y) = (y L ^_) AC- (Flw) A w G A 3 is a onto its image. Further, (Fl^-) A ip^ = A so that the algebraic relations © readily imply that -A(y) A SI = 0, A(Y) A*qQ = 0, i.e. inij4 C A| 7 . Summarising, we can write any (l G A 3 as n = [n]ie[n] 7 ©[n] 2 7 = [6(w A £ + V+)] © [6^- + (X i_V>-) A f + (Alw) A w] © [c(-4w A £ + 3V>+) + (y^-) A £ - (Flo;) A u + /% A £ + 712] (11) 7 for constants d, b, c E R, vectors X, Y E and forms /3g E A|, 712 € A 3 2 . m particular, decomposing = /3 A £ + 7, where and 7 are the uniquely determined 2- and 3-forms in A*^ such that £i_/3,7 = 0, we obtain /3 = (d - 4c)w © (X + F) L ^- ft (12) 7 = (d + 3c)i/) + ©6^_© ((X- y) L w) Aw©7i 2 . (13) Thus j3\ = (a — Ac)oj etc. For later applications, we need for Af £ £ the identities *n AO) = +2X (14) and gn(X^,X^) = 2g n (X,X). (15) We prove (j!4|) along the lines of (|SJ), while (j!5|) uses the transitive and isometric action of SU(3) on (S* 5 . Hence, up to a rotation we may assume that X = \X\e\. Similarly, the transitive action of G2 on 5 6 implies (eA0,cA^) = 4(e,o, (eA*no,eA* n fi) = 3(e,e) (is) for all £ £ A 1 . Furthermore, in conjunction with (j4|) and ([6j we note the useful formulas (r^A^fi.T^A^fi) = 3(t 7 2 ,t 7 2 ), (t 2 AO,t 2 AO) = 4(r 7 2 ,r 7 2 ), (17) (r|AO,r|AO) = 4(r 7 3 ,r 7 3 ) for all vf € A§. 4 The Dirichlet energy functional V In this section we introduce the Dirichlet energy functional and study some basic properties. In particular, we compute its first variation. Definition 4.1 The Dirichlet energy functional V : O^(M) — > R>o is defined by v(n) = l [ {\dn\ 2 gn + \de(n)\ 2 Jvoi n . Remark: Using the L 2 -inner product and integration by parts we may also write z>(n) = ^(ll^llia + llfeon 2 ,) = ±(A a n,n) Ll . Proposition 4.2 (i) The functional T> is invariant under orientation preserving diffeomor- phisms, i.e. V{y*Vt) = X>(0) for all <p € Diff(M)+, O £ O 3 (M). (ii) For A £ M>o, X>(AO) = \aT>(£l), i.e. T> is positively homogeneous. Proof: The first assertion follows directly from = 92* *n V 3 " 1 * f° r V £ Diff(M) + . Secondly, we recall that O = AO £ O^(M) if A > and £ O^(M). Now a G 2 -frame {ej} for O gives the G2-frame {/j = A" 1 / 3 ^} for O. Its dual basis is {/' = A 1 / 3 ^}. Hence, v °l\n = J 1 A ... A / 7 = As voIq, while for the metric g^ induced on A p , we find g\Q = 8 2S 3 we observe that considered as an operator QP(M) -> ft 7 "- p (M), /u...i p ® /ij...^ = A 3 gn . Hence = A 3 \dQ.\ 2 gn . To compute l^ftlg- = |d*Aft Aft|g An 7-2p *\n a p = X 3 * n a p , whence \d* X n A ^lg A n = A ^\ d *^\gn- U Corollary 4.3 The space X of critical points ofT> is acted on by Diff(M) + and is given by ^ = {fie n\{M) I dQ = 0, 5 n Q = 0}, i/ie torsion-free positive 3-forms on M, which are the absolute minima ofD. Proof: The first claim follows from diffeomorphism invariance. Secondly, we can apply Euler's formula for homogeneous functions to get D n V{Q) = |z>(n) = ^(A n ft, 0)^2 > 0. (18) Equality holds precisely if A^ft = 0, i.e. <ift = and ifoft = 0. Hence, if ft is critical, then in particular DqT>(£1) = and therefore ft is torsion-free. ■ Next we compute the first variation of V. To that end we introduce the following piece of notation. Let E be some vector bundle and A : ftj_(M) — > C°°{E) a differential operator. We write Aq for the linearisation of A at ft € ftj_(M) evaluated on some 3-form ft tangent to ft, i.e. A n := D Q A(Q). We illustrate this convention by two examples which will be needed later. Example: (i) Consider the non-linear, homogeneous map e : n 3 + (M) -> n 4 (M), n h- * n ft. Further, for a fixed G2-structure ft € ft^(M) we define the linear, self-adjoint isomorphism pn : n 3 (M) -)■ ft 3 (M), ft ^ § [ft] x + [ft] 7 - [ft] 27 . With the concrete G2~structure in mind we shall simply write p. By Prop. 10.3.5 in |16] . ®n = *npnft- (19) In particular, ft = ft gives 0^ = 4@(ft)/3. (ii) In continuation of the first example we consider the map F : ftj_(iW) — > ft 3 (M) defined by F(ft) = A n n. Then Fq = *nd *n dU + * n d*nd£l + * n d * n dfl - d*ndQ(fl) -d* n d9(ft) (*) = *nd*n rfft — d *q d *q pn^l + terms of lower order in ft = SndQ + d5nPn^ + terms of lower order in ft. Proposition 4.4 We have V n = I ft /\*n{Sndn + Pn d5 n Ft + qn{V n n)) Jm for some quadratic form qn whose coefficients depend smoothly on ft. 9 Proof: As in the previous example, Vq = - I dtl A *ndtl + dfl A (*ndfl + *ndti) + \ f de n a * n d@(n) + de(fi) a (*nde(n) + *nde n ) = / dfi A+od^ + dOn, A*nd9(fi) Jm +- / dn A* u dn + d@(n) A*nd@(n). (20) Now Fq(cZO) : h > i^dO is a bundle endomorphism of A 3 T*M depending on df2 in a fibrewise linear fashion, so we can consider its fibrewise adjoint (T^dO))*. Thus dUMndn = (r n (dn)(n),* n dn) L 2 = (n,(r n (dfi))*(*nd«)) £? . M We deal with the second term of (|20p in a similar manner. The last line is therefore of the form J M Cl A qn(V n Q) with quadratic in the first derivatives of Q, as asserted. On the other hand, Stokes implies / dtl A-kndft + d@ n A*nd@(Q.) = / tt A d*n dtt - A d*n d6(fi) Jm Jm = (&,5adtt + pniddnQ)} whence the assertion. ■ 5 Short time existence Let Q : fl\(M) ->■ S1 3 (M), = -gradP(fi) denote the negative L 2 -gradient of T> in the sense of Definition 4.10 [2], i.e. (Q(Q,),Cl) = —T>q. In view of Proposition 14. 4[ we find Q(O) = -fedtt -pn(^Q^) - 9n (Vn). (21) The goal of this section is to prove the existence part of Theorem 5.1 Given fio £ Q\(M), there exists e > and a smooth family Q(t) G il^(M) /or t G [0, e] such that ^n = Q(n), n(o) = n . (22) Further, if O(t) and are solutions to (|22p . i/ien f2(i) = O'(i) whenever defined. Hence £l(t) is uniquely defined on a maximal time-interval [0, T) /or some < T < oo. Remark: We emphasise that, in contrast to the corresponding results for the Laplacian flow in [B], the initial condition is not assumed to be closed. 10 Definition 5.2 We call the negative gradient flow ofV defined by (|22p the Dirichlet energy flow with initial condition Qq € Q 3 i _(M). We will prove short-time existence and uniqueness by invoking the standard theory of quasi- linear parabolic equations which we briefly recall, see Chapter 4.4.2 [1]. Further useful refer- ences are Chapter 7.8 in jTU] and Chapter 7.1 in |22j . Consider a Riemannian vector bundle (E, (• , ■)). Let Qt : C°°{E) — > C°°{E) be a family of quasilinear, second order differential operators, that is locally, Qt(u)(x) == {a^ (t, x,u,S/u)didjU^ + & Q (i, x, u, Vu)) s a for smooth functions a^ 3 and b a and a local basis {s a } of E. We say that the induced flow equation ^-u = Q t (u), u(0) = u (23) is strongly parabolic at uq if there exists a constant A > such that the linearisation D uo Qq of Qq at no satisfies - {a(D Uo Q )(x,C)v,v) > A|£| 2 M 2 (24) for all (x,£) € TM, £ 7^ 0, and u € E^. Here, the minus sign in fj24[) stems from our definition of the principal symbol. Namely, for a k-th order linear differential operator Q we define a(Q)(x,Ov = ^Q(f k u)(x) for an / G C°°(Af) with /(x) = 0, d-J = £ and u 6 C 00 ^) with = v. Theorem 5.3 If equation (I23p is strongly parabolic at uq, then there exists e > and a smooth family u(t) £ C°°(E), t 6 [0, e] swc/i i/iai ,9 —u = Q t (u), u(0) = u . Further, if u(t) and u'(t) are solutions to (|22p . i/ien ii(i) = u'(i) whenever defined. Next we investigate the operator Q as given in (|2ip . Lemma 5.4 The second order non-linear differential operator Q is quasilinear. Proof: For instance, up to composition with the linear map p whose coefficients depend solely on Q, the second term on the right hand side of (|2ip can be locally written as - d5 n Q ^ (Q)d j (*^(Q)il abc ))d X i ^. Here, *fi c stU v(^) denote the coefficients of *q : Q 3 (M) h-> Q a (M) with respect to local coordinates x\, . . . ,Xf etc. Hence Q is linear in its highest (i.e. second) order derivatives. ■ Lemma 5.5 The principal symbol g{DqQ){x,£) : A 3 T*M — > A 3 T*M of the linearisation of Q at Q £ fij_(M) is giwen fry <7(AiQ)(x,£)n = a n) - TO (e A (^upn(n))). Moreover, the symbol is negative semi- definite. 11 Proof: As the principal symbol involves highest order terms only, we only need to linearise the expression — Sfidfl — pn(d5£iQ) = Q(Q) — terms of lower order in fi. In our convention, a(d)(x,^)Q = i£ A ft and a(5^i)(x, = Hence, from (|19p and the standard symbol calculus we get the asserted symbol. Further, -$n(ff(AiQ)(s,£)n,n) = sn(^An)+pn(^A($Lpn(n))),n) = \^Ah\ 2 n + \^ PQ (h)\ 2 n > o so that a(DaQ)(x, £) is negative semi-definite. ■ Remark: For p> € Diff(M) + we have ip*Q{n) = Q(ip*n) (25) since T> o cp* =T> and (CIq, &i)r2 = (c^ _1 *f2o, ip~ l *Cli) L 2 . Because of this diffeomorphism invariance we cannot expect the principal symbol to be negative definite. Indeed, using the decomposition f2 = /3A£ + 7asin Section O gn(a(DaQ)(x,^)Cl,Clj = implies 7 = 0. From (|13p we deduce d = —3c, 6 = 0, Xq = Yq and 712 = 0. Consequently, one then gets that /3% = — ^Lp(fi) = 0, whence ker a(D u Q)(x, £) = {{vuj + Kif)-) A £ | v € R, V € f 1 } by (HU). In order to apply Theorem 15. 3 1 we use so-called DeTurck's trick which was orginally invented for Ricci flow [8]. Given a family of diffeomorphisms dt<Pt = Xf o ip t induced by a (time- dependent) vector field Xt on M, differentiating (|25p yields the intertwining formula C X (Q(SI)) =D Q Q(£ X Q). (26) Here Cx denotes Lie derivative with respect to X. While the left hand side of (|26p is of first order in X, the right hand side is of third order. Passing to symbol level implies a(D n Q)(x, o a(X ^ C x Q)(x,0 = 0. (27) In this way, we can conceive the symbol of the map q € n%(M) t-> x(n) e c°°{tm) ^ A(n) = £ x(n) ^ e n 3 (M) (28) (where the vector field X(Q) depends non-trivially on the 1— jet of O) as a kind of projection to the kernel of g{Dq j Q). One therefore expects the symbol of the modified operator Q(fi) = Q(Q) + A(fi) (29) to have trivial kernel for a suitably chosen vector field. For a fixed £l € O^(M) we take x n ■. o 3 (M) -> n^M), x fl (n) : = -((y ft n) L n, (30) 12 where we contract and dualise with respect to the metric g^. We think of as a first order, linear differential operator. Subsequently we write and Qq in (|29p to emphasise the choice of £l. To give some motivation concerning the definition of Xq we introduce the operator A^ : C°°(TM) ^n 3 (M), X H> C x &. (31) We consider to be the formal L 2 -adjoint with respect to (• , -) L 2 of A^, i.e. (Afl(X),f2) £ a = (X L (in + d(AY^),r2) L 2 = -(dn,nAX) L 2+(tt,5 a nAX} L 2 whence Aq(H) = —X^(il,) — Q,i_d£l = — Xq(0) + terms of lower order in Q. (32) In analogy with the decomposition of symmetric 2-tensors into a divergence free part and a part tangential to the Difr*(M) + -orbit of some given Riemannian metric we have: Proposition 5.6 For any G n 3 (M) there exist X G C°°(TM) and tl G ft 3 (M) uratft A^(Oo) = such that we have an L^-orthogonal decomposition n = n ® c x n. (33) Proof: Put L = A n A^ : C°°(TM) -»■ C°°(TM). If we can solve Afi(fX) = for some X G C°°(TM), then taking J7o = £l — Cx& yields the desired splitting. Since L is symmetric, such an X exists if and only if A^(r2) G (kerL)- 1 -. But L(Y) = implies A^(y) = 0, whence (A fl (O) , = for all Y G ker L. U Remark: (i) The condition Aq(O) = should be viewed as a gauge-fixing condition, i.e. a choice of a local slice to the Diff{ Af) + -action near Cl. If ft is torsion-free, then A^(f2) = Xq(J1) = (<^r2)i_fl, hence Aq(J1) = if and only if [(5^0] 7 = 0. This is precisely the gauge- fixing condition considered by Joyce in |16] , see also the remark following Corollary 15.91 (ii) The vector field X in the decomposition (j33[) is unique if there are no non-trivial in- finitesimal automorphisms of Cl, i.e. vector fields X such that £x& = 0. This holds for holonomy G2-manifolds as Ricci-flatness implies any Killing field X to be parallel, so that the holonomy is contained in SU(3) unless X = 0. Note further that a generic G2-form has no infinitesimal automorphisms as these are automatically Killing for g^ and a generic metric has no Killing vector fields [9]. An example for a non-generic 3-form is provided by the direct product M = S 1 x CY® of 5 1 with coordinate vector field X = dt and an almost Calabi-Yau manifold (CY e ,uj,ip+) (i.e. a; is a non-degenerate 2-form and is a 3-form of special algebraic type). By ([8|, Q = dt A co + is a G2~form and £x& = 0. Coming back to the mainstream development, we establish strong parabolicity for the flow equation ^0 = Q n (fi), fi(0) = n o - (34) 13 Lemma 5.7 The equation (|34p is strongly parabolic for Qq sufficiently C 2 -close to Q. Proof: Since Xq is linear in f2, we find for the linearisation Aq = oI(Xq((1,)i_Ci) + lower order terms in Cl by virtue of Cartan's formula, whence <7(£> fl A)(x,0n = i^/\{a(X n )(x,0^n) = -£ a ((^0)^0). Assume without loss of generality that = 1. Decomposing f2 = /?A£ + 7as above we deduce from (j5]) cr(L^A)(x, f)fi = -£ A = -3£ A [/3] 7 . Bearing (|4|) in mind, the projection of /3 onto Ay is given by /3i©±(/3 6 +*n(As AO)) Affil((Xo + F )^_ + (Xo + yo)A^). Consequently £ A [/3] 7 = £ A + §(X + YoKV'-) so that using |£A 7 & = |7lfi and gn ^A[f3] 7 ^Af3 + 7 ) = \[(3] 7 \l the computation from Lemma 15.51 implies gn (a(D n Q)(x,on,n) = -| 7 | - 1^(6) ft - 3|Aln - 2 \( x o + *o)mM&. (35) Now =(T0 — /?8 with <7^(ct, = in view of the decomposition in (jlip . while by®, l/S&Hftlfi + lftlfi + lftft- But © gives |/3 6 |^ = |(X + F )^-|, whence - 9n{v(D n Q)(x,0n,n) > (\(3\l + l7lft) = l^lfi by daSJ). ■ Definition 5.8 We ca// t/ie flow associated with (|34p . i/ie Dirichlet-DeTurck flow at f2 with initial condition Qq. If the G2 -form we use to perturb the Dirichlet energy flow is understood we simply speak of the Dirichlet-DeTurck flow with initial condition Qq. For Q,q € il^(M) consider the Dirichlet-DeTurck flow at Q,q with initial condition Qq. By Theorem 15. 3| the flow £l(t) exists on some time interval [0, e]. Let (ft be the family of diffeomorphisms determined by d t cpt = -Xq (fi(i)) oip t , cp = Id M . (36) Then £l(t) = ipl£l(t) is a solution to the Dirichlet energy flow (|22p with same initial condition Qq for # Q(S!). Moreover, the initial condition is satisfied as fi(O) = Id^^o = £Iq- 14 Corollary 5.9 (Existence) For any Qq G J7i(M) there exists an e > suc/i that the Dirichlet energy flow (|22p exists for t G [0, e]. Remark: (i) The idea of DeTurck's trick is to break the diffeomorphism invariance by modifying the flow along the Diff{M ) + -orbits via the additional term Aq . To see this happening in a geometrical way, assume for simplicity that Qq is a torsion-free G2-form. By (|5|), Aq = d(Xn (&)\-£lo) = -3d[(fo f2] 7 . In particular, Aq = if and only if [<fo f2] 7 = 0, for (d[5n ( Ah,n) L 2 =\\[Sn nh\\ 2 L% ■ On the other hand, the tangent space at fio of the Diff(M) + -orbit Oq is given by {Cx^o I X G C°°(TM)} (cf. also Lemma O below). Since S2 S2 () the form f2 — 0,q is perpendicular to Tq Oq if and only if An = 0. (ii) To become strongly parabolic after perturbing with An is a particular feature of the Dirichlet energy flow. In contrast consider the gradient flow of the Hitchin functional 7i restricted to the cohomology class [Qq], cf. (|3|). Upon rescaling T~L, the resulting flow is d —a = 5 no+da (Q + da), a(0) = for a in a suitably small open neig hbourhood of € Q 2 (M) so that (lo + daE £l\(M). The solutions are in 1-1 correspondence with solutions to the Laplacian flow ^-n = f(q) = A n n, n(o) = n . Now a(D^F)(x, £)Q = |£|qO + £ A (£i_([f2]i/3 - 2[J7]2 7 )) and one can easily compute that the kernel is also given by K = ker a(D Q F)(x,0 = {{iu: + T>l^_) A f j v G K, V G ^}. Since by (J27J, the symbol of X 1-4 takes values in if, DeTurck's trick cannot modify the component pr^-± °p(DqF)\ k ± : K 1 - — > K^~. However, the eigenvectors (l\ = P%/\£ and = ^_ in K 1 - give gn(p,i, (t(Dq,F)(x, £)&i) = < and g Q (n 2 ,a(D Q F)(x,0^2) = 4|£|q > respectively. Hence, the linearisation of F = F + A will be indefinite no matter how the vector field X in (|28p is chosen (though the linearisation of F might have trivial kernel). We therefore deal with a heat equation of mixed forwards/backwards type for which short-time existence is in general not expected unless further conditions are imposed. For the Laplacian flow this has recently been achieved in [6] and |25] for closed initial conditions. 6 Uniqueness We now settle the uniqueness part of Theorem 15.11 along the lines of the uniqueness proof for Ricci flow. As shown by Corollary 15.91 a solution to the Dirichlet-DeTurck flow Q(t) with initial con- dition r?o yields a solution to the Dirichlet energy flow f2(i) = <^£0(t) with same initial 15 condition by integrating the time-dependent vector field in (J36|. Conversely, substituting fi(t) by y? i T 1 *r2(t) turns the ordinary differential equation (|36p into the partial differential equation d —(ft = -X Qo (ift U n(t)) o tp u ip = Id M - (37) A curve <pt G Diff(M) + which solves (|37p for a Dirichlet energy flow solution Q(t) with initial condition £Iq yields the Dirichlet-DeTurck flow solution Cl(t) = p^~ u £l(t) with same initial condition. Indeed, let Y% be the time-dependent vector field defined by Y± o ip^ 1 = dttp^ . Then differentiating the constant curve ip^ 1 o (ft(x) = x gives Y t (x) = -d Vt{x) Vt\-Xn {n U m)°Vt{x)) = ^X^ u n(t))(x), (38) where for p € Diff{M) and X € C°°(TM), (tp*X)(x) := d lp -i^p(X(ip~ 1 (x)). As a consequence, we get = Q{vt u m) +c Xno{ip -„ m) ^ u m by (|38p . We can then deduce uniqueness of the Dirirchlet energy flow from uniqueness of the Dirichlet-DeTurck flow, see Corollary 16.31 Remark: Equation (|37p should be considered as an analogue of the harmonic map heat flow introduced by Eells and Sampson |10| . albeit with a time-dependent tension field T g ( t ^ go (<pt)- We can think of T g ( t ^ go (ipt) as a differential operator defined by Riemannian metrics g(t) and go on M, taking a smooth map p : (M, <?(i)) —> (M, go) to a section 7g(t),g ( ( / 9 ) £ C°°{p*TM). We need to prove short-time existence of a solution to (|37p . By and large, we proceed as in the harmonic map heat flow case, cf. jlOJ . Let Pt = Pn(t),n ■ V e Diff{M) + c C°°(M, M) ^ -p,X^n (fi(t)) ° <p e C°°{p*TM). (39) Since Diff(M) + is open in C°°(M, M), a solution to the flow equation d —p t = Pt(tft), = Hm yields the desired solution to (|37p . To get formally in a situation to apply Theorem 15. 3| we first choose an embedding /o : M — > M. n and identify M with its image under fa. In particular, all tensors on M pushed forward to fo(M) will be denoted by the same symbol. Let J\f C K n be a tubular neighbourhood of M which we think of as an open neighbourhood inside the normal bundle ir : I'M — > M, the normal bundle taken with respect to the Euclidean metric on M n . By choosing a fibre metric 16 h and a compatible connection \/ uM on i/M we obtain the induced metric ir*gn + h on Af which we extend to M. n using a partition of unity. In particular, this makes /o an isometry. Similarly, we extend Qq by 7r*f2o to M and subsequently to R n . In this way, the restriction f*0,Q for / in a suitably small open neighbourhood IA C C°°(M, R n ) of embeddings close to /o is still a positive 3-form on M. Consequently, we can extend Pt to an operator P t :U C C°°(M,M n ) ^ C°°(M,R n ), / ^ -df(X f * Qo {n(t))). Lemma 6.1 The operator Pt is a quasilinear, second order differential operator. Proof: Let e\, . . . , e n be the standard basis of W 1 and x , . . . , x 7 be local coordinates on U C M. The components *opgr of *f*n '■ Q 3 (U) i— > ^(U) depend on the components of f*n given by Q 0ja ^d Xl f a d Xm f 1/3 d Xn p . Schematically, so that by the chain rule d*/*n m ^ (a^ a ^(t,a;,a s J«)^ i ^ i /^ + 6 ws (t,a;A I / ai ))^ w ' i for smooth coefficients a;;; and 6.... Applying once more */*n an d contracting the result with f*Clo leads to (★/•n <**/-n mVfVo = {■S i ^(t,x,d Xl f a )d l d j f + b kl (t,x,d Xl f a ))dx k . Finally, dualising and contracting with df = d Xi pdx % <g> e 7 shows that Pt is a quasilinear, second order differential operator. ■ Lemma 6.2 There exists e > and a smooth family of embeddings f(t) 6 C°°(M, IR n ), t € [0, e] such that ^/(t) = Pt (/(*)), /(0) = /o- (40) Furthermore, f{t){M) C /o(M) /or a// 1. Proof: As above let £Y C C°°(M, M n ) be an open neighbourhood of /o. By transversality of /o(M) to the fibres of the normal bundle we may assume that / is an embedding for f (zU with <ff : = 7r o / g Diff(AT) + , shrinking if necessary. We put o~f(x) := /(x) — (fif(x) € ( X }M so that /(x) = <Pf(x) +o~f(x) with . In the following we view af as a section of the pull-back bundle ip*fvM. Observe that af = if and only if f{M) C M. The operator Pt considered as acting on M n -valued functions on M is not elliptic as the computation of its symbol shows below. In order to complement it to an elliptic opera- tor we proceed as follows: For ip £ Diff(M) + consider the connection Laplacian /^P* vM = (y<p* v My<y<p*vM on the pu U-back bundle tp*vM. Here V^ A/ denotes the pull-back con- nection on (p*vM. Its formal adjoint is taken with respect to the metric gn on M and the pull-back metric ip*h on ip*vM. Then, reasoning similarly as above, / i— >■ &?f vM af : U C C°°(M,R n ) C°°(M,ip f uM) C C°°(M,R n ) is a quasilinear, second order differential operator, and so is Pf.Uc C°°(M,R n ) -»• C°°(M,R n ), / ^ P t {f) - A v *f vM a f . 17 We wish to establish short-time existence and uniqueness of the associated flow equation = £(/(*)), /(0) = /o- (41) To compute the linearisation Dj Pq(Y) we take a curve f s dU through /o with Y(x) = -^-f s (x)\ s=0 eT x R n ^R n . as We write YH(x) and Y ± (x) for the projections of Y(x) to T X M and v x M. By design of the extension of f^o to R n (cf. our convention above), <p s := tv o f s G Diff(M) + satisfies /gfio = Vs^o- Furthermore, one clearly has ^ s (x)| s= o = ^ ll (^) and -^a s (x)\ s=0 = Y ± ( y x) (42) for cj s (x) := f s {x) - (p 8 (x) G v^^M. First we compute the linearisation of Pq. Using the naturality of the vector field Xq , i.e. <p*X 9 *n (£l(t)) = X no ((p- u n{t)) for (p G Diff(M) + , we obtain -j^Po(fs)\s=0 = —-j^dfs(X<p*Q Q (£lo))\ s =0 = -±df s (^ l x no (<p; u n ))\ s=0 . For Y tangent to M, i.e. Y{x) = Y'l(x) for all x G M, we may in view of (|42p assume that f s = (p s G Diff(M) + for all s and we get — (5n d(Y l_$1q))i_Qq + terms of lower order in Y. For Y perpendicular to M, i.e. Y(x) = Y^(x) for all x G M, we may assume that ip s = Id for all s, again using (142 p . Hence ^Po(fs)\s=o = -^-df s XQ (n ) = dY(X no (n )), which is of lower order in Y and hence does not contribute to the symbol. For general Y = Y' + Y 1 - we therefore find Df Po(Y) = — (5n d(Y^ i_Qo))l-Qo + terms of lower order in Y. For the linearisation of / >— > A tp f vM Gf we again assume Y to be tangent to M first. For a curve f s = ip s G Diff(M) + as above we have in particular that a s = for all s. Hence 4-A (p * sVM cr s \ s= Q = 0. For Y perpendicular to M we have <j s (x) = f s (%) — x for a curve f s with 9? s = Id as above. Hence £ A^ uM a s \ s=0 = A uM Y. For general Y = yll + Y- 1 we get £> /o (/ h-> A ¥ '/ l/M (7 / )(Y) = A^Y^. It follows that the symbol of the linearised operator Df Po is ff (£> /o p )(z,e)^ = -(£-(£ a (yii L no)))ufio - \Z\lvY 1 - IS To check strong parabolicity we assume |£|n = 1 and write Y" = a£ + Yq, a G R, Yq G £~ and Qq = uj A £ + ^M-- Then (7T*5no + h)(a(D f0 P )(x,OY,Y) = g no ((^(Z A (Y^n )))^ ,Y\\) + \Y~ 2 I ft gn {n , (£ L (e a (yii L n ))) a y") + y^ 5 n (y |l ^o,£L(eA(yiiLOo))) + |y ± i^ 3n (aw + *ol-^+ + (yoLw) A £, aw + y L^ 4 sia^ + iyo^+l^ + ^l 2 , 3|a| 2 + 2|y | 2 + |y ± |^ Theorem 15.31 applies once again to yield short-time existence and uniqueness of (|4ip . Last we show that a solution f(t) to (|4ip satisfies f{t)(M) C M for all t. As in this case 0"/(t) is just the zero section of (p(t)*vM, we obtain the desired solution to (140)) , We proceed as in the harmonic map heat flow case, see for instance Part IV in |13| ; Consider the bundle endomorphism r : vM — >• vM which is multiplication by —1 in each fibre. We claim that Pt o r = dr o P t . This is clear for Pt since (r o f)*Qo = f*Qo- Furthermore, (p ro f = <ff and a ro f = r o of, such that by linearity of the connection Laplacian one has A<°f uM <j rof = r o A^ M a f = dr o A^f vM a f , where the last equality follows by viewing a section of vM as a vertical vector field on the total space of that bundle. This proves our claim and implies that for a solution f(t) to ([41 p the composition r o f(t) is again a solution with initial condition Jq. Now if f(t)(M) were not contained in M for some t, then r o / would yield a second, different solution with same initial condition, contradicting uniqueness. ■ A solution f(t) eU to pj) yields a solution p t = / -1 o /(i) G Diff(M) + to ([37} for a given Dirichlet energy flow solution Q(t). From there, uniqueness easily follows: Corollary 6.3 Suppose that Q(t) and Q'(t) are two solutions to ([22]) fort € [0, e], e > 0. If 0,(0) =fl = fi'(0), then fi(t) = JV(t) /or a// t € [0,e]. Proof: Solving for ([37p with and J)'(i) gives two flows (ft and <^ which without loss of generality we assume to be defined on [0, e]. By design f2(i) = ip^Q(t) and Q'(t) = ip' t *Q'(t) define a solution to (j34|) at f^o- Uniqueness of the Dirichlet-DeTurck flow implies Q(t) = Q'(t). Hence (ft and ip' t are solutions of the ordinary differential equation ^ t = -X Qo (n(t))o^ t . (43) By uniqueness of the solution to ()43j) . we conclude (ft = ip' t , whence Q(t) = Q'(t). ■ 7 The second variation of V In this section we compute the second variation of D at some fixed Cl G Af = Q _1 (0) (cf. Corollary 14 . 3 1) . Further, we show that X is a Frechet manifold whose tangent space at Cl is 19 precisely hevD^D. All metric quantities, projections on G2-invariant modules etc. will be taken with respect to this Cl. For a given vector bundle E —¥ M we denote by W k ' 2 (E) the space of sections whose local components have square integrable derivatives up to order k. The associated Sobolev norm will be written || • ||y(/fc,2. Further, we simply write L 2 (E) for W 0,2 (E). More generally, we consider the Hilbert manifolds W k,2 (^) for a fibre bundle £ — > M in order to deal with non- linear differential operators. The integer k will be chosen appropriately when required, but at any rate big enough so that all sections involved are at least of class C° and the corresponding function spaces W k ' 2 (M, 1R) are Banach algebras under pointwise multiplication. According to the Sobolev embedding theorem we therefore need k > 11/2 for Q to extend to a smooth map Q k : W k ' 2 (A s + T*M) W k ~ 2 ' 2 (A 3 T* M) . Lemma 7.1 The space X k := Q^" 1 (0) is a smooth Banach manifold whose tangent space at Cl G X is given by T n X k = {Oe W k ' 2 (A 3 T*M) I d(l = 0, d<d n = 0}. Proof: We have X k = {0 G W k ' 2 (A 3 + T*M) \dtt = 0, = 0}, so it remains to show that the extension of n -mg n 3 + (M) >-> (dfi,de(n)) g ^ 4 (m) x n 5 (M) to W k ' 2 (A 3 + T*M) has as a regular value. By the first example of Section [4J (Nk)n — (dfl, d*p(fi)) . Since the range of (d p ) k : W k ' 2 (APT*M) -> W k ~ l > 2 {A p+l T* M) is closed, (B p+l ) k ~ l = im(cZp)fc is a Banach space. Hence for a given (da,dr) G (£? 4 ) fc_1 x (l? 5 ) fc_1 , we need Q G Q 3 (M) such that <if2 = tier and d-kp((l) = dr. The Hodge decomposition theorem of [n] q gives H(n q )@da q @8p q , where d 9 G W fc+1 ' 2 (A 2 T*M), $ q G H/ fc+1 > 2 (A 4 T*Af) for 9 = 1, 7, 27, and H denotes projection on the space of harmonic forms. Similar decompositions hold for a and r. Taking j3 q = [j3 a ] q yields dtt= d5$ q = d5fi a = da. <je{l,7,27} On the other hand, d*p(Cl) = d-kiy^dai + da? — dc^) = d5fi T = dr, provided we put d\ = 3 * [/3 r ]i/4, aj = *[/3 T ]7 an d ^27 = *[A-]27- Consequently, (A^)q is surjective, whence the result by the Banach space implicit function theorem. ■ Consider the closed linear subspace Vj = {O G W fc - 2 (A 3 T*M) | [<5f2] 7 = 0} C A -1 (°) (cf. (|3"2"|) ) and let s n :=v£nx k . Note that S s C Q^(0), whence S n C ST2 3 (M) for Q n is a quasilinear, elliptic operator by Proposition 15.71 Proposition 7.2 Near Q, the space Sq is a smooth submanifold of X k . Its tangent space at 0, is naturally isomorphic with the space of ti -harmonic 3-forms. In particular, dimS^ = &3- 20 Remark: In Section [8] we actually show that near f2, (0) coincides with Sq (cf. Corol- lary EH]). Before we can prove Proposition 17.21 we need a technical result first. Consider the Hilbert manifold Diff(M)Q +1 obtained from completion of the identity component of Diff(M) with respect to || ■ ||^fc+i,2. Then Diff(M) k+1 is a smooth Banach manifold and a topological group which acts continuously on X k via pull-back. We denote by 1^ the subgroup of Diff(M) k+1 fixing n, i.e. ip*Q = for all (p G 1^; in particular g = ip*g which by [2^ implies (p G Diff(M) since any ip G DiS(M) k ) +1 is of class C 1 . Hence 1^ is contained in Diff(M) Q . Let denote the Diff(M)o +1 -orbit through in X k and put G k ^ := Diff(M)o +1 /In- As in Section 5 of [9] one can prove that G^ +1 endowed with the quotient topology is a smooth manifold. Lemma 7.3 If k > 11/2, then [ip] G i— > (p*Q G is an injective immersion with closed image. In particular, is a closed, smooth submanifold of X k with tangent space T^O^ = {C x & | X G W k+1 ' 2 {TM)}. Proof: We can argue as in Ebin's proof of the corresponding result for the moduli space of Riemannian metrics, cf. Section 6 in [9]. The only remaining issue to check is the injectivity of the symbol of A* : C°°(TM) — > Q 3 (M) defined in (|3T|) (this ensures that the extension to a map W k+1 > 2 (TM) -> W k > 2 (A 3 T* M) has closed range). Indeed, a(\*)(x,0(v) =^A(«4 (44) and this vanishes if and only if vi_£l is of the form rj A £ for some rj G ^(M). But v i_f2 G A^, so that in this case = [rj A £]i4 = (2r/ A £ — *(?7 A £ A £l))/3. Since on the right hand side the first term contains £ while the second does not, this can only hold if r] A £ = 0, i.e. if r\ = or equivalently, v = 0. ■ Proof: (of Proposition 17. 2p Since 0^ C X k , the tangent space TqO 1 ^ is contained in T^X k . By extending Proposition 15.61 to Sobolev spaces, we deduce T^X k + TqV^ = W k,2 (A 3 T*M). Hence, the intersection is transversal near 0, so that Sq is a smooth sub- manifold of W k ' 2 (A 3 T*M) in a neighbourhood of f2. The tangent space at f2 is r fl s fi = n r n x k = {ne n\M) \dn = o, de^ = o, [sn] 7 = o}. Next, the map Cl G TqSq i— )■ [O] G H S (M, R) is an isomorphism. For injectivity, assume 0, = drj so that [(W^ = 0. By Lemma 10.3.2 in |16] . this implies <5[<ir/]i = and 5[<ir/]7 = 0, whence 8[drf\2j = for d-kp{drf) = 0. Consequently, Vl = drj is A-harmonic which is impossible unless 0, = 0. For surjectivity, recall that the projections on irreducible G2- components AqT*M commute with the Hodge Laplacian A since 0, is torsion-free (cf. for instance Theorem 3.5.3 in |16]). Hence, a p-form is A-harmonic if and only if its irreducible components in Qq(M) are A-harmonic. So, given a cohomology class c = [ti] G H S (M, M) with unique A-harmonic representative O, we have d*p(fL) = and thus Q G T^Sq- B 21 In particular, we deduce in conjunction with Proposition 15.61 that T a X k = T a O^®T a S a . (45) Next consider the map on a suitable neighbourhood Sq of f2 in Sq. Following |16] $ fc is well-defined: Indeed, near 0, we can linearise the action of 1^ on via the exponential map exp^ : TqSq = H 3 (M, M) —> Sq induced by (• ,-)l2. The linearised action must be trivial as any element in Iq is homotopic to the identity. Hence Iq acts trivially on 5q close to Cl. The Banach space inverse function theorem and (|45p imply that is a diffeomorphism onto its image near ([LIm];^)- Hence, shrinking Sq possibly further, we have shown that Sq defines a slice for the Diff(M)£ +1 -action on X near f2. The same statement holds for the C°°-topology instead of the Sobolev topologies, whence the Corollary 7.4 (i) The space X is a Frechet manifold whose tangent space at 0, is given by T n x = {!ie n 3 (M) | dn = o, de n = o}. (ii) (Joyce) The space of torsion-free G2 -structures modulo diffeomorphisms isotopic to the identity is a smooth manifold of dimension 03 = dim H 5 (M, R). We can now prove the central result of this section. Proposition 7.5 Let Q E X. (i) We have D}p((l,(l)= I (\d(l\ 2 + \d<d n \ 2 ) vol n . J M In particular, the second variation of T) at Q, is a positive semi-definite bilinear form with kevD 2 n V = T u X. (ii) The linearisation Lq := DqQq of Qq is a symmetric, non-positive and elliptic operator given by L n fl = -5dQ - pddptl - 3d[5tt} 7 . More precisely, writing £1 = fCl © ★(d A fi) ® 7, we have - L n n = aq + |f44/ • o + *(44« A n) + 4?4 7 7 - 1;44 7 7 • ^ - f 4?4/ ( 46 ) /or i/ie G2 -differential operators d\ introduced in Section [3J in particular (L^n,n) L 2 = -\\dh\\ 2 Li - \\5ph\\ 2 L2 _-3\\[m] 7 \\ 2 L2 _ and kerL n = T n S n . a a S2 f2 Proof: (i) We compute the second variation of 2? at a critical point f2 by differentiating equation (|20p once more. Thus we obtain DlV(tl,n)=[ dtl A -kdtl + d&Q A *dQ n = A (|dr2| 2 + |rfe^| 2 ) wo/ > 0, 22 for the remaining terms involve either d£l or dQ(Q). But these terms vanish since f2 is a n critical point and hence torsion-free. Furthermore, the kernel D^T> is precisely T^X by Corollary \7A (ii) Since D^D((l,(l) = (Hess^Pfi, f2)^2 and DqQ = — Hess^P, it follows from (i) that D^Q = —5d£l — pddpVL. On the other hand, using again that $7 is torsion-free, we get D n A n (ti) = d(x n (nyn). Bearing ([5]) in mind, we obtain A^ = —3d[5£l]j, whence L n (tt) = -5dtt - pdSptl - 3d[5n} 7 . This operator is clearly symmetric and non-positive. Ellipticity was proven in Lemma 15.71 To compute (|46p we start from ([7J and use Tables [1] and [2] Then Sdtl = i(444/ + 44 7 7)-^©*((44a+|44« + i4 d 7 7 7) ^7^7/ + dl 7 dja + ^dl 7 dj 7 j + dljdm d5n = |(344/ -44 7 7) -^©*(i(44a- 144 7 7) Afi) from which (|46p follows by applying Table [3] d\ 7 d\f — d 7 21 d 7 7 a + |4r4 7 7 + 47417 Let D denote the associated Hodge-Dirac operator with respect to the metric induced by f2 € X, i.e. Bh = dh + 50. In particular, D is symmetric and D 2 = A. In view of longtime existence to be established in the next section we note the following corollary. Corollary 7.6 (Garding inequality) For all Cl € S1 3 (M), (-L n n,h) L2 > \\Bfi\\l 2 . In particular, we have (-^n^i^L 2 > CII^IIvk 1 . 2 ~~ II^IIl 2 for some constant C independent of CI. Proof: Writing £1 as in Proposition 17.51 the first inequality follows from (-L n n,n)v = ||DO||| 2 + f ||4/||| 2 + 4||4a||| 2 + ||4 7 7||l 2 -|(4 7 7,4/)L 2 (47) > \\bq\\i 2 + ^ii^/iii, + \\ d fj\\i 2 - §(n4/iii 2 + ii4 7 7ii| 2 ) lL 2 - > ||Df2| |2 The second inequality is just the elliptic estimate H^H^i^ < C 1 (||r2|| 2 2 + ||Dii|| 2 2 ) for some constant C" 1 . ■ 23 8 Stability We continue to fix a torsion-free G2-form fi € X. From now on, we let W k ' 2 and W+ 2 be shorthand for the Sobolev spaces W k ' 2 (A 3 T*M) and W k ' 2 (h? + T*M) with respect to the metric g^. The induced norm will be denoted by || • ||iyfc,2. Again, k is an integer strictly greater than 11/2 which for simplicity we assume to be odd (this avoids using fractional Sobolev spaces below). In particular, W k ' 2 embeds continuously into C 2 . The goal of this section is to prove the subsequent stability theorem. Theorem 8.1 (Stability) Let Ct 6 fij_(M) be a torsion-free G2 -/orm. For all e > there exists some 6 = 5(e) > such that for any fio with \\£Iq — £l\\ W k,2 < 5, the Dirichlet-DeTurck flow fit at Q with initial condition Qq 1. (longtime existence) exists for all t £ [0, oo), 2. (a priori estimate) satisfies the estimate \\£lt — f2||jyfc,2 < e for all t £ [0, oo), and 3. (convergence) converges with respect to the W k,2 -norm to a torsion-free G2 -form —¥ 00. Since the Dirichlet energy flow exists as long as the Dirichlet-DeTurck flow exists we imme- diately obtain: Corollary 8.2 Let Q € fi^(M) be a torsion-free G2 -form. For initial conditions suffi- ciently C°° -close to 0, the Dirichlet energy flow exists for all times and converges modulo diffeomorphisms to a torsion-free G2 -form floo, i-e. there exists a family of diffeomorphisms (fit £ Diff(AL) + such that (f^t converges to fioo with respect to the C°° -topology. The proof of Theorem 18 . 1 1 will be subdivided into a sequence of intermediate steps. First, we tackle existence of the Dirichlet-DeTurck flow together with the a priori estimate on arbitrary, but finite time intervals for initial conditions sufficiently close to fi. Here we use the Banach space inverse function theorem, following the approach of Huisken and Polden |15] for geometric evolution equations for hypersurfaces. Let < T < 00. If 7T : M X [0, T] — )■ M denotes projection onto the first factor, let C°°(M x [0, T], 7r*A 3 ) denote the space of smooth sections of A 3 pulled back to M x [0, T\. For any non-negative integer s we define the Hilbert space F s [0,T] as the completion of C°°(M x [0,T],7r*A 3 ) with respect to the inner product given by {n 1 ,Q 2 ) V s [0yT] = V / e- 2 *(^fii,^fi2) M/ 2 (s -i),2o!i. In particular, ||fi||^ j 0T j = e~ 2t \\(l\\ 2 L2 dt. These spaces are also known as anisotropic or parabolic Sobolev spaces, where a time derivative has the weight of two space derivatives. The density e~ 2t is introduced for technical reasons, see below. Similarly, we can consider the Hilbert manifold V1[0, T] consisting of sections of 7r*A 3 i_ of class V s . Define the map F:Vl%T] xV'-^O,!], F(fi) = (fi , d t ti t - Q fl (n*)). (48) As usual, restricting to the boundary is tantamount to invoking a trace theorem, which in this context is stating that the trace map ft 1— > fig is continuous from V s [0, T] to W 2s ~ 1,2 , see 24 |15] , We wish to show that F is a local diffeomorphism near Q. Consider the linearisation at £l of the map (|48p . namely D^F : F s [0, T] — > W ' x V [0,T], D n F{fl) = (O ,PO), (49) where PQ := d t tl t - L^fl t . Let H be the completion of C°°(M x [0, T], vr*A 3 ) with respect to the inner product <ni,n 2 >H= [ T e- 2t {fi 1 ,n 2 ) W i,2dt+ C \- 2t (d t fii,d t (i>i) V idt. Jo Jo Note that the quadratic form Cl i— > (— L^fi, 0)^2 is defined on W 1 ' 2 in view of equation (|4 so that we can say that f!eH satisfies Pft = 6 for 6 G V°[0,T] wea£%, if (d t n, *} v o [0 ,t] + / e 2 *(-L n ft, tf) L2 dt = *) y o [0 , T ] JO holds for all ^ € Cq°(M x (0, T), 7r*A 3 ), the space of smooth sections of 7r*A 3 which vanish near the boundary M x {0,T}. We first show that for <3> £ V [0, T] there exists a unique weak solution in H to the equation £> fl F(fi) = (0,i). (50) Following [15], we use a refined version of the Lax-Milgram lemma, cf. Lemma 7.8 and Theorem 7.9 in |15] or Theorem 16 in Chapter 10 of |12| . As explained in |15j . the main point is to check coercivity of the bilinear form A(n 1 ,n 2 )= [ e~ 2t {d t ti u dA 2 ) L 2dt+ [ e- 2t {-L^ti u d t ti 2 ) L 2dt Jo Jo on C£°(M x (0,T),vr*A 3 ) C H, i.e. to establish an estimate of the form A(tl,Cl) > C||fi|||[ for some positive constant C. This will be a consequence of the Garding inequality for the operator — Lq of Corollary 17. 61 First note that f2 is a solution to PCl = <E> if and only if e~*0 is a solution to (P + l)e _ *f2 = e _ *$. Hence by replacing — by — + 1 we may assume that — Lq satisfies a strict Garding inequality of the form {—LqCi,Ci) L 2 > C||0||^i ;2 with C as in Corollary 17.61 Since dt(e~ 2t (— Lq&, tl)ip.)dt = by assumption, we obtain f e- 2t (-L n n,d t n) L2 dt= [ e~ 2t (-L n n,n) L 2dt > c [ e- 2t ||n||^ li2 dt, Jo Jo Jo whence A(n,(l)>C [ e- 2t \\n\\ 2 vl , 2 dt+ [ e- 2t \\d t Cl\\l 2 dt > C||f2||H, Jo Jo establishing coercivity and the existence of a weak solution. In order to improve regularity of this weak solution one needs the following estimate: Lemma 8.3 (Huisken Polden) Let s > 0. If 17 G H is a weak solution to the equation D n F(n) = (n 0) *) with tl G W 2s+1 > 2 and 6 G V s [0,T], then fi G V s+1 [0,T] and there exists a constant C = C(Cl, s) > such that \\&\\v s + 1 [0,T} — C(ll^o|liV2 s +i,2 + ll^llyafcT])- 25 Proof: See Lemma 7.13 in [15]. ■ For later use we state and prove the ensuing interior estimate: Corollary 8.4 (Interior estimate) Let s > 0. For all S > there exists a constant C = C(S, fi, s) >0 such that for fi G V s+1 [0, T] one has 11^ II ya+i^T] — C(ll^llyo[o,T] + ll-f^lly s [o,T])- Proof: Since <1> = Pfi is of class V s , Lemma 18.31 gives |l^l|y-s+l[0,T] — C(£l, s)(j\Clo\\yy2s+l,2 + 1 1 Pfi 1 1 [ 0jT] ) ■ Put Q! := ip ■ fi, where (p : [0, T] — > M is a smooth cut-off function satisfying (pit) = for t close to and tp(t) = 1 for t G [e, T]. Then Pfi' = <9j</? ■ fi + ■ Pfi and Slg = 0, whence ll^llv+Me.r] ^ ll^'lly s + 1 [o,T] < C(fi, s)a ||c^ • fi + 93 • Pf2||ys[ 0| T] < C(fi,s) s (sup te[0iT] |%>(*)| • ||Ji||y S [ 0) T] + ||Pn||y.[o,r|)- Induction on s yields the result. ■ Mutatis mutandis one proves along the lines of Theorem 7.14 in |15] : Theorem 8.5 (Huisken Polden) The map D^F in (|49p is a Banach space isomorphism. As a consequence we obtain existence of the Dirichlet-DeTurck flow together with the a priori estimate on a finite time interval [0, T] for initial conditions (depending a priori on T) sufficiently close to fi: Corollary 8.6 For all e > and < T < 00, there exists 5 = 5(fi, e, T) > such that for fio £ W k ' 2 with \\£lo — £l\\ W k,2 < 5, the Dirichlet-DeTurck flow fi^ a t fi with initial condition fio exists and satisfies ||fi( — fi||jy fc . 2 < 6 f or a ^ * ^ [0, T]. Proof: Since k is odd, s := {k + l)/2 is an integer. Let fi also denote the pull-back to M x [0,T]. Then fi G V s [0,T] and F(fi) = (fi,0). By virtue of the previous Lemma 1831 and the Banach space inverse function theorem, the map F in (|48p is a local diffeomorphism near fi. The trace maps fi G V s [0,T] 1 — >• € W k ' 2 are continuous with a uniform bound on their norms since t varies within a compact interval. Hence there exists C > such that 1 1 fit — fiUjyfc^ < C||fi — fi||ys[o,T] holds for all t G [0, T\. For suitably chosen 5 > 0, the condition ||fio — fi||iy*;,2 < 5 implies that (fio,0) is close enough to (fi,0) to ensure that fi = F -1 (fio, 0) satisfies ||fi - fi||ys[ 0j T] < e /C- ■ Remark: Using Theorem 18.51 it is also possible to give an alternative proof of short-time existence for the Dirichlet-DeTurck flow, cf. |15] for details. As in Section [7] let Sq C Q^^O) denote a suitably chosen slice around fi. For a positive 3-form fi' close to fi we write fij = fi' + oj[. Let now 2(3 and Then we can recast the flow equation into ^t t = L^u' t + R^{u' t ). (51) The basic idea is that the behaviour of the Dirichlet-DeTurck flow should be dominated by the linear term L^,. To obtain precise results, we need to control the remainder term R^,. In order to analyse Rq, we introduce the following notation: Let E\,E2,E and F be tensor bundles over M equipped with the bundle metrics induced by Cl. Denote by V the covariant derivative associated with £l. Let Q be a further positive 3-form of class C 2 and let f2' £ 5jj. Put (J = Q — Q' and assume that ||u/||c2 < ei and — £1\\q2 < e 2 - For s 6 T(M,E) we will generically write ®is for sections of the form A(s) € T(M, F), where A 6 T(M, E* © F) is a section of class C l possibly depending on to' and £l' such that ||^4.||c?i < C(ei, e 2 ). Using this convention, we have for instance ®k ®is = ®kS for k < I and, by the product rule for V, V(©/s) = ®i_ 1 s + ©;Vs. Similarly, for a section B £ T(M, El ® £* ® F) with ||-B||^i < C(ei,e 2 ) and Sj € T(M,Ei) we write generically si ®i S2 for the section B(si,S2) € r(M, -F). For instance we have (®l So) ®k (®hSi) = si ® k s 2 , and © fc (si ©; s 2 ) = si ©fc S2 for k < I, Iq, l\. In all cases we set © := ©o- As we will differentiate at most twice we take I to be smaller or equal than two even if we have C^-control for higher I. Example: Since the exterior differential d is obtained by anti-symmetrising the covariant derivative V, we have du' = ® 2 Vuj' . Hence *Qidoj' = © 2 Vu/ and we get d*Q,duj' = © 2 V(©2Vw') = © 2 (©iVu/ + © 2 vV) = §iVw' + © 2 v V. The following result gives a rough description of the structure of the remainder term, which however will turn out to be sufficient for our purposes. Lemma 8.7 Let Q, be a positive 3-form of class C 2 and £l' G Sq. Let further u' = 0, — fV. Assume that \\lo'\\ C 2 < e\ and ||fi' — Q\\q2 < e 2 . Then R n , (J) = J ® J + J ® Vw' + J ® V V + Vw' © Vu/. (52) In particular, there exists C = C(ei,e2) > such that \Rq>{u')\ < eiC(|u/| + |Vo/|). 27 Remark: In particular we absorbed into © any term in (|52p of order strictly higher than two in u/. Proof: Recall that Qn(fi) = Q(0) + A n (fi) where Q(n) = -(A n n + ^[^qO]i - 2[dfe^]27 + ?fi(vn)), projections being taken with respect to CI, and with X(f2) = — ((5^0)l0. In the following we will calculate the difference between Qq and its linearisation at Cl' term by term. First write fl = fV+u/ and observe that mapping a positive 3-form £1 to the corresponding Hodge operator *q gives rise to a fibre-preserving smooth map -k : A^T*M — > End(A*T*M). Hence there exists a fibrewise linear map ^4q/(u/) : A 3 T*M -> End(A*T*M), depending smoothly on 0' and u/, such that * n = * n , + j4q, (u/)u/; in particular (_Dfy*)(u/) = Aqi(Q)uj'. Using df2' = 0, we get 5fidO, = -kfid-kfiduj = *n' d *n' du>' +*n,d(Aw(uj')L>')du)' + {A^{u)')uj')d-k^ du>' +{A^')J)d{A^,{J)J)dJ. Subtracting the linear term in lj', we get SfidQ — *&id *Qr du> = ■Xa' d ( A ci'( u} 'W) du >' + (An,(ui')u>')d*w dJ + {A^{J)uj')d{A^(uj')J)dJ = J © Vw' + u/ © V 2 u/ + Vw' © W. Similarly using 5qiCi' = 0, we get dSntl = -d-knd-knO, -d(A iil (u,'W)d* (i , J - d{A^')J)d{A^{J)J)V! -d{A- n ,(J)J)d{A- n \J)J)J . Again, subtracting the linear term in u/, we get dSuQ + d-kQ, d <J + d ^(Aq/ (0)u/)fi' = -d * fl/ d((A n , (a/) - A n , (0))a/)fV - d ★ft, d(% (w>'y - d(^,(u/y)d(^VV)^ = u'®u' + J® Vuj' + uj'® VV + Vu/® Vw'. This takes care of the first three terms in Q (note that the linear projections onto irreducible components are absorbed into ©). Furthermore, the term coming from the quadratic form contributes a remainder term of type Vu/ © Vu/. Finally, in order to deal with the term Aq(J1), we observe that X(Cl') = since CI' € Sq which implies [<5qJY]7=0. Hence X(Cl) = X{J) and Zxiu)')^' = UJ ' ® Vu/ + J © V 2 u/ + Vw' © Vu;'. This finishes the proof. ■ 28 In the following we need some standard results from perturbation theory of linear operators. This is summarised in the following statement: Lemma 8.8 For all e > there exists 5 = 5(e) > such that if — fl\\ W k,2 < 5, then (-Lftu, uj) L 2 > (1 - e)(-Lna;, u)u» - e\\u\\ 2 L 2 for alluj € TV 2 ' 2 . Proof: We apply Theorem 9.1 in |24| to the operators T := —gLq and V := Lq — Lq,. Without loss of generality we may assume that Lq, is symmetric; otherwise we replace it by the symmetric operator \(Lq, +L^,), where * denotes the formal adjoint taken with respect to Cl. Then by elliptic regularity for the operator T the estimate H^^Hl 2 < a||cj||^2 + 6||Tw||x,2 holds for arbitrarily small a, b > if Cl' is sufficiently close to Cl with respect to the W k ' 2 - norm. In particular, V will be T— bounded with T-bound less than 1. Since T is non-negative we obtain with T + V = —Lq, + (1 — c)Lq that - L n , + (1 - e)L n > -e or equivalently (-L qi uj,uj) L 2 > (1 - e)(-L Q u,u}) L 2 - e\\uj\\ 2 L 2 for all to G W 2 ' 2 and fl' sufficiently close to Cl. ■ The family (Lqi)qi^ s is a smooth family of elliptic operators, and hence gives rise to a smooth family of Fredholm operators Lq, : W 2 ' 2 — > L 2 . Since Sq is a smooth (finite- dimensional) manifold and Sq C (0), we have Tq,Sq C kerLjy. Furthermore TqSq = kerLfj by Proposition 17.51 Usual Fredholm theory implies that Tq,Sq = ker Lq, for fl' sufficiently close to Cl, and hence the kernels and ranges of Lq, form smooth vector bundles (the latter infinite-dimensional). Interpreting (ker Lq,) 1 - as the fibre at Cl' of the normal bundle of Sq with respect to the L 2 -metric, we obtain the following statement: Lemma 8.9 (Orthogonal projection) There exists e > such that if \\Q — Cl\\ W k,2 < e, then there exists a unique Cl' E Sq such that cj' = O — &' G (ker Lq,) , the orthogonal complement being taken inside L 2 . Remark: In the situation of Lemma 18.91 we obtain using the continuity of the embedding W k ' 2 ^-?> C 2 that there exist constants 61,62 > such that ||0 — 0||^fe,2 < e implies that ||w'||c<2 < ex and \\Q,' — &\\c 2 < e 2 with e% , €2 — > as e — > 0. Proposition 8.10 For all k > there exists e > such that if ||0 — ^^,2 < e, then \\Rq,{u')\\ L 2 < k\\Lq,oj'\\ L 2 with £l' £ Sq and uj' = — Q! as in Lemma \8.9l 29 Proof: Elliptic regularity implies that L fr : (kerL n ,) X n P^ 2 ' 2 -> imL n , C L 2 is a Banach space isomorphism, the orthogonal complement being taken inside L 2 . In particular, there exists C > such that ||u;||^2,2 < C||Lq/u;||j / 2 for all oj 6 (ker^/)-* - !"!^ 2 ' 2 . This constant C can be chosen uniform in Q', since L^/ is a smooth family of elliptic operators parametrised by a finite-dimensional, hence locally compact manifold. Using Lemma [8. 71 we get \\R n ,(uj')\\ L 2 < exC'\\cj'\\ W 2,2 < e 1 CC'\\L n uj'\\ L 2 with ei, e2 > as in the remark following Lemma [8. 91 and some constant C" = C"(ei, £2) > 0. For ei 1, this constant C' can in fact be chosen independent of ei in the above estimate, which implies the result. ■ As a consequence, we obtain that locally Q^^O) coincides with 5^, more precisely: Corollary 8.11 There exists e > such that if ||0 — $111^,2 < e, then Qq(O) = implies that fi € Sq. Proof: Let 0,' € Sq be chosen according to Lemma 18. 9| i.e. such that oj' = fi — Cl' sat- isfies oj' £ (kerLQ/) -1 -. We write Qn(fi) = LqiOj' + Rq/(oj'). Now Qn(^) = implies that I ^^/(a;') 1 1 £,2 = HLq/Cj'11^2, which in view of Proposition 18.101 (with e.g. k = g) yields Rq,(lj') = Lq/Oj' = 0. Since oj' was chosen orthogonal to the kernel of Lq/, we obtain oj' = 0, hence fi £ Sq. ■ Let fit be a Dirichlet-DeTurck flow solution on [0, T] and let Qt '■= Qft(fif). Then J^fit = Qt and Qt satisfies the linearised flow equation |<2* = (53) with = D^Qq. This is a linear parabolic equation with time-dependent coefficients. We view the operator as a (non-symmetric) perturbation of the symmetric operator Lq, to which in particular Lemma 18.81 applies. Lemma 8.12 (L 2 almost orthogonality of Q t ) For all n > there exists e > such that if \\Qt — 0,\\ W h,2 < e for all t € [0, T], then I (Q<,^)l 2 I < K ll<2i||z, 2 IMlL 2 for all oj € ker Lq. Proof: Let £ be chosen according to Lemma [8. 9 1 i.e. such that oj' t = Oj — Oj satisfies oj' t € (ker L^J" 1 ". We write Qt = L^ t Wt + -RQt( a; t)- First we observe that due to the symmetry of Lq its range is L 2 -orthogonal to its kernel. Since the ranges of the operators Lq, form a smooth vector bundle for Cl' € 5q, this implies that they are L 2 -almost orthogonal to ker Lq in the above sense, in particular I ( l qM,u)l 2 \ < K \\ L nM\\L 2 \\^\\L 2 for all oj £ kerL^, if e > is chosen sufficiently small. Proposition 18. 101 then shows that I (RnA^'t)^)^ < II^^IMMIl 2 < K||LQ t ^|| i2 ||w|| L 2 30 and hence I (Qt,u) L 2\ < I (L^ t uj' t ,uj) L 2\ + I {R^ t {J t ),u) L 2\ < 2k \\L^ t uj' t \\ L2 \\uj\\ L 2 < 2k(1 - «) _1 ||Qt||i2||a;|| i 2. for all u € kerL^, again if e > is chosen sufficiently small. This proves the result. ■ Lemma 8.13 (L 2 — exponential decay of Q t ) There exist e > and A = A(fj) > such that if ||fii — n|||yit,a < e for all t € [0,T], then \\Qt\\h<e- xt \\Q \\h. for all t € [0,T]. Proof: Using equation (|53p we obtain Let Ai > be the first (positive) eigenvalue of — Lq. Then one has (— LqUJ, uj)l 2 > for all oj G (keiL^)- 1 -, and hence, using Lemma |8.12| that ~ ~ 3Ai ~ {-LciQt,Qt)L 2 > -^-IIQtllia for e > sufficiently small. Using Lemma 18.81 one gets {-L^Q t ,Q t ) L2 > (1 - e ')±f\\Q t \\l 2 -e'\\Q t \\ 2 L2 > f\\Q t \\h if e, e' > are chosen sufficiently small. Finally we get j t \\\Qt\\h < -X\\Qt\\h with A := Ai/2 and e > sufficiently small. Now Gronwall's lemma implies the result. ■ Parabolic regularity theory now yields higher derivative estimates: Proposition 8.14 (W k ' 2 — exponential decay of Q t ) There exist e > andC,\ > such that if — 0,\\ W k,2 < e for all t £ [0, T], then \\Qt\\ 2 wk ,2 < Ce- Xt for all t G [0,T]. Proof: If e > and A > are chosen according to Lemma I8.13| then for < to < ti < T we get -At / 1 WQrWhdr < \\Q \\l 2 f 1 e~ Xr dr < C G e J to Jtn 'to J to for Co := ||Qo||£,2/A. It is easy to see that the estimate of Corollary 18.41 remains true for Pt := dt — (with a constant C s = C s (e, S, O) > 0) if e > is sufficiently small. In particular we get for s := (k — l)/2 that ll<2llys+i[i +<5,t 1 ] < C r a||Q|lv'0[to,ti] — ^ / IIQrll^ u _dr to since PtQt = 0. Combining these two estimates and using the trace theorem as in the proof of Corollary 18.61 yields the result. ■ 31 We will now describe the choice of the constants: 1. Choose e > such that (a) \\Qq — Q\\\yk,2 < e implies that the Dirichlet— DeTurck flow at with initial condition Qq exists on [0,1]. This is possible using Lemma 18.61 (b) \\U t - n\\ W k,2 < e for t € [0, T] yields HQ t \\W k ' 2 < Ce -\t on [0, T] for any T < T max , where T max is the maximal life-time of the flow. This is possible using Proposition 18.141 2. Choose T > such that Ce~ xt dt < e/2 with C and A as above. 3. Choose 5 > such that ||fio — f2||yyfc,2 < S implies that the Dirichlet-DeTurck flow at f2 with initial condition exists on [0, T] with \\ttt — &\\w k - 2 < e /2- This again is possible using Lemma 18.61 The proof of Theorem 18 . 1 1 may then be finished as follows: Let the initial condition Oo be given satisfying Let T max > T such that [0,T max ) is the maximal time interval on which the Dirichlet- DeTurck flow with initial condition Qq exists. Suppose now that T max < oo. For T < t < T m ax one has n t = n T + d_ f2 T dr and hence, as = Qt, W k ' 2 < < Q T dr , w k < 2 \\Q T \\w k > 2 dT Using this for to = max{T, T„ mind, we obtain /•oo / Ce~ Xr dT < e/2. Jt 1/2} and bearing the assumption — 0||pi/fe,2 < e/2 in n t0 < \\n t n T\\W k < 2 + \\Sh 0.\ w k,2 < e. Therefore the Dirichlet-DeTurck flow can be continued at least up to time to+1 > T max +l/2, a contradiction. Hence T max = oo and we have established longtime-existence together with the estimate ||fit Finally we set n\ w k 2 < e for all t E [0, oo). n + ot fit dt. 32 Using Proposition 18. 14l we see that this integral converges, i.e. that Of converges to JIqo with respect to the W k,2 -norm. Furthermore, since Qq : W k ' 2 — > W k ~ 2 ' 2 is continuous (even differentiable) , we obtain = lim Q t = Qq(9,oo), t— too in other wor ds, € Q^(0). Since locally Q^(0) = according to Corollary 18.111 we obtain 0^ E Sq. In particular, is a torsion-free G2-form. This finishes the proof of Theorem 18.11 ■ Acknowledgements The authors would like to thank Richard Bamler, Natasa Sesum, Miles Simon and Jan Swoboda for explaining their related work to us. A Appendix: G2— differential operators The G2-differential operators d\ we use in the text are given as follows: d\ : Qi ' d\f = = d\f = df , 4rf = dl : fii — > Q g d\ dja d 7 27 a 5 n a = \ *nd( *n (a A Q) A Si) *u(da A *n0) = -~ *n (fa (a A 0) A O) [da] 14 [d *^ (a A *n^)]27 = [fa(a A n)] 27 dj 4 : df/3 714, 4 4 /3 -*n ([#An) [d/9]27 df : fii -»• ? [&7]l4 *si(fe7A*s)n) ^fi[d7] 2 7 References [1] T. AUBIN, Some Nonlinear Problems in Riemannian Geometry, Springer, Berlin, 1997. [2] A. BESSE, Einstein manifolds, Springer, Berlin, 1987. [3] E. BONAN, Sur les varietes Riemanniennes a groupe d'holonomie G 2 ou Spin(7), C. R. Acad. Sci. Paris 262 (1966), 127-129. 33 [4] R. Bryant, Metrics with exceptional holonomy, Ann. Math. 126 (1987), 525-576. [5] R. BRYANT, Some remarks on G 2 - structures, in: Gokova Geometry/Topology Confer- ence (GGT), pp. 75-109, Gokova (2006). [6] R. BRYANT AND F. Xu, Laplacian Flow for Closed G 2 - Structures: Short Time Behav- ior, Preprint 2011, arXiv:1101.2004. [7] S. CHIOSSI AND S. Salamon, The intrinsic torsion o/SU(3) andG 2 structures, World science publishing (2002), 115-133. [8] D. DeTurck, Deforming metrics in the direction of their Ricci tensors, J. Diff. Geom. 28 (1983), 157-162. [9] D. EBIN, The moduli space of Riemannian metrics, in: Global Analysis, vol. 15 of Proceedings of Symposia in Pure Mathematics, pp. 11-40, AMS (1968). [10] J. EELLS AND J. SAMPSON, Harmonic maps of Riemannian manifolds, Am. J. Math. 86 no. 1 (1964), 109-160. [11] M. Fernandez, A. Gray, Riemannian manifolds with structure group G2, Ann. Mat. Pura Appl. 132 (1982), 19-45. [12] A. FRIEDMAN, Partial differential equations of parabolic type, Prentice Hall, Englewood Cliffs, 1964. [13] R. HAMILTON, Harmonic maps of manifolds with boundary, Lecture Notes in Mathe- matics 471, Berlin 1975. [14] N. HlTCHIN, Stable forms and special metrics, Contemp. Math. 288 (2001), 70-89. [15] G. HuiSKEN AND A. POLDEN, Geometric evolution equations for hyper surf aces, in: S. Hildebrandt, M. Struwe (ed.), Calculus of Variations and Geometric Evolution Prob- lems LNM 1713, pp. 45-84, Springer, Berlin, 1999. [16] D. JOYCE, Compact manifolds with special holonomy, OUP, Oxford, 2000. [17] S. Karigiannis, Flows of G 2 -structures, Q. J. Math. 60 (2009), no. 4, 487aA§-522. [18] S. KARIGIANNIS, Some notes on G2 and Spin(7) geometry, in: Recent advances in geometric analysis, Adv. Lect. Math. (ALM), 11, pp. 129-146, Int. Press, Somerville, MA, 2010. [19] O. Ladyzhenskaya, V. Solonnikov and N. Uraltseva, Linear and Quasilinear Parabolic Equations, Nauka, Moscow, 1967. [20] U. SEMMELMANN AND G. WEINGART, The Weitzenbock machine, Comp. Math. 146 no. 2 (2010), 507-540. [21] N. SESUM, Linear and dynamical stability of Ricci-flat metrics, Duke Math. Jour. 133 no. 1 (2006). [22] M. TAYLOR, Pseudodifferential Operators and Nonlinear PDE, Birkhauser, Basel, 1991. [23] M. TAYLOR, Existence and regularity of isometries, Trans. Amer. Math. Soc. 358 no. 6 (2006), 2415-2423. 34 [24] J. WEIDMANN, Linear Operators in Hilbert Spaces, Springer, Berlin, 1980. [25] F. Xu AND R. Ye, Existence, Convergence and Limit Map of the Laplacian Flow, Preprint 2009, arXiv:0912.0074. H. WEISS: Mathematisches Institut der Universitat Miinchen, Theresienstrafie 39, D-80333 Miinchen, F.R.G. e-mail: weiss@math.lmu.de F. Witt: Mathematisches Institut der Universitat Minister, Einsteinstrafie 62, D-48149 Miinster, F.R.G. e-mail: f rederik . witt@uni-muenster . de 35